2011 International Conference on Advanced Technologies for Communications (ATC 2011)

A tool for analysis and classification of sleep stages

Le Quoe Khai, Truong Quang Dang Khoa+, and Vo Van Toi

Biomedical Engineering Department
International University
Ho Chi Minh City, Viet Nam
[email protected]

Abstract The scoring of sleep-wake stages is usually performed by a

highly trained human expert on the basis of an epoch-by-epoch
Scoring sleep stages is a critical process in assessing several sleep visual interpretation of the PSG signals according to a set of
studies and slumber disorders. Sleep is classified in two major R&K rules, and recently updated by AASM 2007 rules. PSG
states: non-rapid-eye-movement (non-REM) sleep and REM records uses a fixed 30s epoch duration and allows for the
sleep. Non-REM sleep comprises stages Nt, N2 and N3. We
recognition of different sleep and wake stages. Manual scoring
develop a tool for automatic scoring the stages of sleep following
relies on visual extraction of specific features in EEG channels
the rules of 2007 AASM (American Academy of Sleep Medicine).
(usually C3-A2 and C4-A1 of the International 10/20 system
The study propose the algorithm to classify based on some
[4]), two channels of EOG and one channel of chin EMG [3].
different characteristics of each stage, due to using a device of
It is known that such manual work of sleep staging is very
polysomnography (PSG) in order to collect the signals of
complex and onerous. This process is time consuming
Electroencephalography (EEG), Electro-oculography (EOG) and
Electromyography (EMG). Methods of analysis are Fast Fourier
(typically 1-2 hours per subject), costly (millions of
Transform (FFT), Candidate of REM (CREM) and Digital Signal Vietnamese dong per recording) and prone to inter and intra
Filters (High pass, Low pass, Notch Filter). PSG signals were scorer variability [5]. Although performed by trained
recorded continuously overnight in 5 healthy volunteer students personnel, the reliability and coherence of manual sleep scoring
(19 - 25 years old, 4 males and 1 female). PSG data are analyzed are usually unsatisfactory because of subjective judgment and
in 30 second epochs (data windows) in offline mode. The main human error [6]. In order to overcome the problems associated
result of analysis and classification is a hypnogram which were with manual scoring several researchers have proposed
compared with those obtained by an experienced human scorer. automatic sleep scoring system.

Keywords: automatic scoring, sleep stages, polysomnography, With the advancement in digital signal processing
AASM, hypnogram techniques, several other used frequency spectral analysis [7],
neural network analysis [8], multidimensional scaling and
wavelets techniques or expert system approaches [9] to develop
I. INTRODUCTION
automatic sleep staging system. We thus describe the
developed staging process combine many methods to detect
Sleep was described as a succession of five repeating stages features of each sleep stages. For each data, we apply many
by Rechtschaffen and Kale (R&K) in 1968: the rapid eye tool to filter, analyse both time domain and frequency domain
movement (REM) stage and four nonrapid eye movement to collect the valuable information to calculating. After
(NREM) stages, SI, S2, S3, S4 [I]. It has not been changed classification and post-processing, the results was been
until 2007 the American Academy of Sleep Medicine updated checked again to against some errors by analysing in single
part of it. S3 and S4 were grouped into one stage and signed epoch. A description of the algorithms corresponding to these
N3 stage [2]. Thus following the AASM 2007 rules, NREM different processes is also provided.
sleep has three major stages: NI, N2, N3 replace for SI, S2, S3
or S4, respectively.
II. METERlALS AND METHODS
Polysomnography (PSG) is a psycho-physiological method
for the assessment of sleep and wake states. It is based on the
concurrent recording of brain electroencephalography (EEG), A. Experimental Data
chin electromyography (EMG) and electro-oculography (EOG)
signals collected in human individuals using non-invasive Before each experiment, subject was explained for the
surface electrodes. PSG allows for the description of different purpose and procedures of the study. Then they were asked to
sleep-wake states, which may exhibit abnormal qualitative and fill out a questionnaire which was kept confidential and
quantitative changes with clinical conditions and included patient's identification. Informed consent and health
environmental situations. It is the golden standard in the information inquiry form that had been filled out by the
diagnosis of sleep disorders [3]. volunteers was also obtained. Before they sleep, we did not use
any kind of medicine, stimulant or sedation. The whole night

978-1-4577-1207-4/11/$26.00 ©2011 IEEE 307

polysomnography was recorded using clinical PSG equipment for many channels (Table I) to observe the appearance types of
(Alice 5, Respironics, Phillips [10], Fig l.) at the Sleep wave in each epoch. We use an different threshold value of
Laboratory of Biomedical Engineering Department of average amplitude of voltage to separate the bad signal come
International University. During sessions, subjects were lying from electrode placement.
on a comfortable bed within a dark and sound-isolated room.
After the experiment, they did not complain anything about the
TABLE!. THRESHOLDS APPL Y FOR SIGNAL AFTER USING FFT
experiment in terms of their health and condition of Lab-room.
Four male and one female subject voluntarily participated in Raw data + FIT Threshold of
Appear Disappear
this study. All of these subjects were healthy and free of Power (uV')

neurological and psychological signs and symptoms. The data Alpha (8-13Hz) 12 1 0
of each subject, including EEG, EOG and EMG signals. In this Beta (>13Hz) 4 1 0
study, one channel of EEG were analyzed from the C3-A2
derivation, two channels of EOG include Left EOG and Right Theta (4-7Hz) 21 1 0
EOG signals with a sampling rate of 100Hz, one channel of Delta (1-3Hz) 43 1 0
Chin EMG with 200Hz in sampling rate.
Slow waves (1-3Hz) 800 1 0
EEG cup-electrodes were attached onto the scalp of the
EOG «0.5Hz) 2500 1 0
subjects according to the international 10-20 system for
electrodes placement. EOG and EMG electrodes were attached
onto the skin [3], Fig 2. The patient preparation, and To separate the rapid-eye-movement s (REMs) from slow­
instrumental setup were done according to AASM guidelines eyes-movements (SEMs), prior to any detection, the two
[2] and Sleep Technician Guide of Alice 5 device [11]. channels of EOG data are filtered using a fourth-order digital
Butterworth bandpass filter with cut offs at 1 and 5Hz to yield
Sleep-wake stages were scored according to conventional
the filtered EOG data . This effectively minimizes false
criteria using a fixed epoch duration of 30s. Each epoch was
detections due to SEMs as REMs as well as any high frequency
classified exclusively amongst six possible stages: wakefulness
noise [13]. We detect the Candidate of REM in each epoch
(W), transitional sleep (N1), shallow sleep (N2), deep sleep
which is the characteristic of REM appearance [14]. Detection
(N3), REM sleep (R) and movement time (MT).
of rapid eye movement is based upon correlation coefficients
between two EOG derivations and deflection between left and
right EOG.
EGelectrode

30 seconds I epoch
C3M2 (EEG). LOC + ROC (EoG), EMG

Fig.l: PSG Alice 5 Fig.2: Polysomnography for sleep study

B. Methods

Automatic sleep stages classification is effectuated in the

hierarchical manner, presented in Fig 3. The first step in
process of data analysis is apply digital signal filters for each
channel of EEG, EOG and EMG signal. The filters have been
used in this case include three types: 50Hz of Notch filter to
remove the noise of power source voltage fluctuations,
Lowpass filter and Highpass filter to have a adaptive
bandwidth for bio-signals. The specification of filters for each
channel is 0.3HzHighpass and 35Hz Lowpass for EEG and
EEG signals, 10HzHighpass and 100Hz Lowpass for EMG
signal [2].

We then used the Fast Fourier Transform (FFT) to No

transmute the signal from time domain to frequency domain. It
N2
is easy to detect and analyze the feature waves of EEG waves
in power-frequency scale, for example Alpha, Beta, Theta,
Delta waves which is define base on frequency range [12].
Beside the spectrum of EOG signal allow we to distinguish the Fig 3: A procedure of our software using to analyse and score sleep stages

rapid or slow movement of eyes. Applying different thresholds

308
 • Detection of Body movements or Artifact signals m. RESULTS

Movement Time (MT) is a separate stage assigned in case The system used for all the computations, including the
of EEG, EOG or EMG signals obscured in more than half the algorithm of automatic sleep stage classification, is complete
epoch by muscle tension or amplifier blocking artifacts source code base on Matlab computing language. We develop
associated with movement of the subject. Movement artifacts this user interface for anyone use our software easier and
are usually reflected in a big increase of the amplitude of EEG, permit they manually adjusting the thresholds value for some
EOG and EMG signals. The values of amplitude signal are specific case, for example: with human that have a sleep apnea
compared to thresholds, determined from the training data: ± or older people, threshold to detect appearance of Alpha wave
300 flV. The tonic EMG activity differs across the sleep stages must be higher than another people in normal. By this way, a
and time of sleep. It almost always reaches its lowest level doctor can check correctly scored epoch following traditional
during stage REM-this fact is used in the scoring of sleep method that he or she has more experiments. (Fig 4)
stages. To determine relative level of tone EMG signal, for
each 30-s epochs of filtered EMG signal. This procedure was ) Full Sirep Sl� r-l-!5Cl
adopted to reject the burst-type high activities of EMG EE GWalies Thresholcl ofP'ower
Pnks owtr
Thrutll:lld
occurred also in stage REM. Thresholds for the value of tone �hO(8-IlHz)S-- r=-
-l _____
- � 12 Epo<n .. n I
EMG, detecting stage REM, were determined on the basis of eoto(I3-30Hz)£J � • Epoch""", SOl
The"'(4.7Hz)� I .1 ,.
the data collected to train the system. These thresholds are O""(1.�Hzj� C.I 0 [ R..PIot I
computed for each sleep recording as the average EMG tone of _ ......S(02.o1Hz) ·r [ '1 !IX)
all the epochs (with the additional constraint of not exceeding
10 flV).

• Assigning Stages to Epochs

For each 30-s epoch is tested for muscle artifacts in EEG,

EOG or EMG derivation, which can indicate the Movement
Time (MT). Relevant parameter reflecting increase in 0.­

amplitude was calculated as described in Section "Detection of

_I o Epo<h(Vlow) I
Body movements or Artifact signals". If at least one of the
� 0S<",,,, oSloge

analyzed derivations C3A2, Left or Right EOG, EMG

exceeded a corresponding threshold 300 flV, the epoch is Fig. 4: Screenshot from a main window of our software, the full
scored as MT. procedure built up all parts of this software is described in this paper. Another
one window - epoch (view) allow user easy to show the signal in both time
In the next step, the algorithm detects Alpha wave, which is domain and frequency domain for 3 types of channel: EEG, EOG, EMG with
a feature of wakefulness by applying band-pass frequency for each epoch.
filtered channel C3-A2 of EEG signal from 8 to 13 Hz.
Similarly, we apply frequency band to detect appearance of
The main result of all analysis process in our software is the
Beta, Theta and Delta waves. Delta waves or slow wave is a
hypnogram, the diagram of sleep stages per epoch. Processing
goal standard to distinguish the shallow sleep and deep sleep.
is time latency about 2-4 minutes per subject compare with 1-2
Stage REM is scored in case of tone of EMG below 10 flV, in
hours per subject in classical scoring methods using visual
presence of stage 1. In the absence of theta waves-as stage
interpretation of scorer. However, that results have been
Wake. This is the alternative way of scoring stage Wake for
compared with the manual hypnogram based on experiment of
subjects who do not present clear alpha activity.
a doctor with sleep and sleep disorder as speciality. Total
concordance of the proposed automatic detection of sleep
stages with hypnograms by human experts, scored for these
• Smoothing the Hypnogram epochs, is greater more than 75%. Table IT.A show a high
agreement between hypnogram of our software analyse and
doctor scored. Moreover, the result of our software expressing
According AASM 2007, the epoch-by-epoch approach
that we can overcome some errors of the hypnogram is created
presented in the previous section does not imply that each
by Alice 5 software analyse. Alice 5 Sleepware Software v.2.8
epoch is considered in isolation. We explicitly implement the
is a software supporting combine with PSG Alice 5 device, this
following, generally accepted smoothing rules: the "3-minute
rule" for stage N2. Smoothing rules of replacing consecutive software is normally install in a computer be use to recorded

epoch of R N2 R, N2 N1 N2, W R W with R R R, N2 N2 N2, the signal of PSG. It has a tool in the same function of our
software to automatic scoring sleep stages, but results of our
W W W, respectively [15]. At the end of the procedure,
software and that software have big difference. These
hypnogram is reviewed for elimination of I-epoch stages
differences in scoring of analyzed data are summarized by
(except MT).
statistical measures in Table Il.B, H.C.

309
 TABLE II. A: STATlSTIS OF THE CONCORDANCE RESULTS OF ALlCE 5 single channel of EEG, EMG and two channel of EOG. All
AND OUR SOFTWARE
major stages are plotted in hypnogram is call Macro Sleep
MT W REM Nt N2 N3 Architecture (MSA). Estimation of MSA is a critical process in
n 11 369 99 6 309 13 assessing several sleep disorders such as Obstructive Sleep
Apnea (OSA), Periodic Leg Movements (PLMS) and Upper
ALICE 5 %n 1.4 45.7 12.3 0.74 38.3 1.61
Air Respiratory Syndrome (UARS) [16]. However, one of the
n 15 206 129 48 345 64
most difficult challenges for analyzing structure of sleep is
Proposal %n 1.7 25.5 16.0 6.0 42.8 7.93
detection of the micro parts or events occur in a period of time
Concordance n 0 182 88 0 193 8
---
which is much less than 30 seconds. So in the near future, we
of ALICE 5 %n 0.0 38.6 18.7 0.00 41 1.7 will add more tools to detect micro structures of sleep such as
& Proposal % 58.36 micro-arousals, K-complexes, sleep spindles.

B: STATlSTIS OF THE CONCORDANCE RESULTS OF OUR SOFTWARE AND

DOCTOR SCORE ACKNOWLEDGMENT

This work is partly supported by International University of

MT W REM Nt N2 N3
Vietnam National Universities in Ho Chi Minh City. We would
n 15 146 131 20 342 153
like to thank Center for Community Health Care (CHAC) for
Doctor %n 1.86 18.1 16.2 2.48 42.4 19.0
helpful clinical suggestions and advices. We would thank Dr.
n 15 206 129 48 345 64 Dang Vu Thien Thanh for expert comments and reference
Proposal %n 1.86 25.5 16.0 5.95 42.8 7.93 materials. Our BME volunteer students is an important
Concordance n 15 142 113 0 302 64 contribution in the study.
of Doctor & %n 100 22.3 17.8 0.00 47.5 10.1
Proposal
% 78.81 REFERENCES

[1] Rechtschaflen A, Kales A "A manual of standardized terminology,

C: STATlSTIS OF THE CONCORDANCE RESULTS OF ALlCE 5 AND DOCTOR SCORE techniques, and scoring system for sleep stages of human subjects".
Washington: Washington Public Health Service, US Government
MT W REM Nt N2 N3 Printing Office; 1968.
n 15 146 131 20 342 153 [2] "The AASM Manual for the Scoring of Sleep and Associated Events -
-
r---- --- -------

Doctor %n l.86 18.1 16.2 2.48 42.4 19.0 Rules, Terminology and Technical Specifications", American Academy

n 11 369 99 6 309 13 of Sleep Medicine,2007

[3] Aliki Minaritzoglou, Emmanouel Vagiakis. "Polysomnography: Recent

ALICE 5 %n l.36 45.7 12.3 0.74 38.3 l.61
data on Procedure and Analysis". PNEUMON Number 4,Vol. 21.
Concordance n 0146 99 0 196 12
-
r---- - --- ------
[4] Kelly ACarden. "Recording Sleep: The Electrodes, 10/20 Recording
of Doctor & %n 0.00 32.2 21.9 0.00 43.3 2.6 System,and Sleep System Specifications".Sleep Med Clin 4,pp.333-41
ALICE 5 % 56.13 [5] Collop NA, "Scoring variability between polysomnography
technologists in different sleep laboratories". Sleep Med,3(1): pp. 43-47

[6] Norman, R. G "Interobserver agreement among sleep scorers from

Taking the stages separately, mean concordance between diflerent centers in a large dataset". Sleep Journal,2000,23:pp.901-908.
visual scoring and automated analysis was 20% for sleep stage [7] Van Hese P, Philips W, and et ai, "Automatic Detection of Sleep Stages
Nl, 88% for stage N2, 50% for stage N3, 86% for REM stage, using EEG". 23rd Annual Conference - IEEEIEMBS Oct.25-28, 2001,
and 95% and 100% for stages Wake and Movement Time, Istanbul,TURKEY,2001. 2: pp. 1944-1947.
respectively. These concordances exhibit a clearly different [8] B.Kim, K.P., "Automatic sleep stage scoring system using genetic
pattern across the stages for the system of Alice 5 software - algorithms and neural network". Engineering in Medicine and Biology
Society,Proceedings of the 22nd Annual International Conference of the
expert and system of our software - expert cases. Therefore,
IEEE,2000. 2: p. 849 - 850.
system of Alice 5 software - expert concordance for Movement
[9] M. Qianli, et aI., "Sleep-stage Characterization by Nonlinear EEG
Time, stage 1 and Wake is higher than system of our software­
Analysis using Wavelet-based Multifractal Formalism". Engineering in
expert agreement. Finally, the concordance for stage 2 proves Medicine and Biology Society, Proceedings of the 27th Annual
similar agreements more than 85%. It is a good result because a International Conference of the IEEE,2005: p. 4526- 4529.
normal adult spends approximately 50 percent of total [10] Philips, Respironics, Alice 5,Alice Sleepware v.2.8 software manual.
sleep time in stage 2. [II] Philips,Respironics,Alice 5,Sleep Technican Guide.
[12] William O. Tatum, IV, DO, Aatif M. Husain, MD, Selim R. Benbadis,
MD, Peter W. Kaplan, MS, FRCP, "Handbook of EEG Interpretation",
TV. CONCLUSIONS
Demos Medical Publishing,2007
A difficult challenge is staging criteria defined in AASM [13] P. Shokrollahi et al. "A Method for Quantifying Sleep Eye Movements
2007 leave a significant margin for subjective interpretation; That Reflects Medication Eflects". IFMBE Proceedings (25) pp.1411-14
therefore, hypnograms constructed for the same recording, rely [14] Rajeev Agarwal, Tomoka Takeuchi, Suzie Laroche, Jean Gotman.
on doctor experiments. Although the concordance of our "Detection of Rapid-Eye Movements in Sleep Studies", IEEE
Ransactions on Biomedical Engineering. Vol. 52,No. 8,August,2005
software analyse and doctor scored is high, some different
staged epoch show that we need more and more data for [15] Virkkala, 1., Hasan, 1., Varri, A, Himanen, S.-L., & Muller, K,
"Automatic sleep stage classification using two channel electro­
training our system to find a best threshold value.
oculography. Journal of Neuroscience Methods,2007,166,pp. 109-115.

In this paper we present a tool for automatic scoring [16] Kryger, M.H., T. Roth, and w.e. Dement, "Principles and Practice of
Sleep Medicine". 3 ed. 2000:W. B. Saunders . 1336.
Movement Time, Wakefulness and four stages of sleep using

310